Application of data-driven machine learning to predict propranolol and trimethoprim removal using a managed aquifer recharge system

Abstract

Owing to their persistent nature, pharmaceutical products (PPs) are emerging as potent water pollutants. Here, experimental and data-driven modeling, specifically multilayer perceptron (MLP) neural networking and gene expression programming (GEP), was employed to predict the removal of the most common antihypertensive and antibiotic drugs, namely propranolol and trimethoprim, from reclaimed water (RW) through a managed aquifer recharge system (MARS). The characteristics of RW and soil used as the column medium, including operating time (days); pH; dissolved organic carbon; electrical conductivity; and concentration of nitrogen dioxide, nitrate, sulfate, ferrous, chloride, and manganese, were included as the input parameters and removal of the selected PPs as the model output. A dataset was created through an experimental study conducted over a year of continuous operation of MARS to predict the removal of the selected PPs. MLP and GEP models were developed for one of the selected PPs and tested for the other to determine model reliability. The developed models were assessed using statistical performance matrices. The experimental results showed over 80% propranolol and trimethoprim removal from RW through MARS. The proposed GEP predictive models for propranolol and trimethoprim removal showed higher accuracy (R2 = 0.91 and 0.87, respectively) than the MLP models (R2 = 0.827 and 0.756, respectively). Therefore, the proposed GEP models provide better predictions and mathematical relationships for future studies. Thus, data-driven machine learning models can predict the removal of specific PPs from RW through MARS and minimize the experimental workload.